Recently, Google DeepMind has integrated its SynthID technology into the Gemini model and released it as an open-source project. SynthID for Text employs sophisticated processes to intervene in the text generation of large language models (LLM).
SynthID slightly adjusts the probability scores of the next token predicted by the LLM, provided that it does not affect the quality, accuracy, or creativity of the output. This process is repeated for all generated text, forming a watermark.
According to Google DeepMind, this technology can be applied to as few as three sentences. For longer texts, the watermark becomes more robust and accurate. Although the method is suitable for various languages, it has some weaknesses when editing AI-generated text.
Google DeepMind has integrated SynthID into the Gemini application and website to add watermarks and identify generated text. The technology is also available as an open-source project on GitHub, Google Responsible Generative AI Toolkit, and Hugging Face.
Google DeepMind claims that SynthID outperforms existing text watermarking systems. Previously, Google DeepMind had introduced SynthID for images, speech, and music.